Keras is a high-level neural network API that runs on top of TensorFlow. Instead of writing 500 lines of low-level TensorFlow ops, you define a neural network in 50 lines: stack layers (Dense, Conv2D, LSTM), compile with an optimizer and loss, and train. Keras abstracts away the complexity of gradient computation, backpropagation, batch processing, GPU management. It supports feedforward networks, CNNs (image), RNNs/LSTMs (sequences), transformers (attention), and custom architectures. Models can be saved, deployed to production, or converted to mobile (TFLite).